tasks/ml/manifests/ML22.yaml

# ============================================================================
# T22 — Active learning query strategies for logistic regression on small pools
# ----------------------------------------------------------------------------
# Unlike paper_replication's P01-P07, the "synthesis" here frames a research
# QUESTION rather than a known paper's method. The model must design the
# experiment (conditions, metrics, datasets) — we only commit to what a
# competent study of this topic would include and what the rubric expects.
# ============================================================================

id: ML22
title: "Active learning query strategies for logistic regression on small tabular pools"
arxiv_id: null
venue: "ARC-Bench 2026"
paper_asset: null

# The "synthesis" plays the role of the upstream briefing: research question,
# background, why the question matters, what "a reasonable experiment" looks
# like. It deliberately does NOT pre-specify a single method to reproduce.
synthesis: |
  Pool-based active learning is attractive when labels are expensive but a
  modest unlabeled pool is available. For linear probabilistic models such as
  logistic regression, classic query policies include uncertainty sampling,
  margin sampling, query-by-committee (QBC), and expected-error reduction
  style lookahead approximations. These methods are often taught as broadly
  useful, but on small tabular datasets their gains can be inconsistent due to
  model misspecification, class imbalance, and high variance from tiny initial
  labeled sets.

  A useful CPU-bounded benchmark should compare these query strategies under a
  fixed annotation budget and shared learner, rather than mixing in model
  architecture changes. The core signal is label efficiency: how quickly test
  performance improves as labels are acquired. Because final accuracy alone can
  hide early-budget differences, the study should include both
  accuracy-at-budget and a budget-curve summary metric such as area under the
  learning curve.

  The experiment should therefore run multiple seeds, use at least two
  sklearn-resident tabular classification datasets, and enforce identical
  initialization, batch size, and stopping budget across strategies. A random
  query baseline is essential to determine whether sophisticated policies
  provide real value beyond chance selection. A passive full-data reference is
  also useful for contextualizing attainable performance ceilings under the
  same logistic-regression family.

  Practical constraints matter: expected-error reduction can be expensive if
  naively recomputing retrains for every candidate. A tractable variant can
  evaluate a capped candidate subset per round and use one-step hypothetical
  label outcomes, preserving the spirit of expected future loss minimization
  while staying within single-core runtime limits.

  *Do uncertainty, margin, QBC, and approximate expected-error reduction deliver better label efficiency than random sampling for logistic regression on small tabular pools under a fixed annotation budget?*

hypotheses:
  - id: H1
    statement: "At 30% labeling budget, at least 2 of {uncertainty_sampling, margin_sampling, qbc, expected_error_reduction} achieve higher mean test accuracy than random_sampling by ≥1.5 percentage points, averaged over ≥5 seeds."
    measurable: true
  - id: H2
    statement: "expected_error_reduction attains the highest mean area_under_learning_curve (AULC) on at least 1 of 3 datasets, and its AULC is not lower than random_sampling on any evaluated dataset."
    measurable: true
  - id: H3
    statement: "qbc outperforms uncertainty_sampling in mean test accuracy at early budget (10% labels) on at least 2 of 3 datasets."
    measurable: true

experiment_design:
  research_question: "Which active learning query strategy provides the best label efficiency for logistic regression on small tabular pool-based classification tasks under fixed budgets?"
  conditions:
    - name: "random_sampling"
      description: "Baseline pool-based active learning that queries unlabeled points uniformly at random each round."
    - name: "uncertainty_sampling"
      description: "Queries samples with highest predictive entropy from current logistic regression model."
    - name: "margin_sampling"
      description: "Queries samples with smallest top-2 class probability margin (binary: |p-0.5|)."
    - name: "qbc"
      description: "Query-by-committee with 5 bootstrapped logistic regression committee members; selects by vote entropy/disagreement."
    - name: "expected_error_reduction"
      description: "Approximate one-step expected-error reduction over a capped candidate subset per round using hypothetical labels and expected future log-loss reduction."
  baselines:
    - "random_sampling is the primary active-learning baseline"
    - "passive_full_data logistic regression reference trained once on all labels for context"
  metrics:
    - name: "accuracy_at_budget"
      direction: "maximize"
      description: "Test accuracy at 30% labeled budget, averaged over seeds."
    - name: "aulc"
      direction: "maximize"
      description: "Area under the test-accuracy-vs-labeled-fraction curve from initial seed set to 30% budget."
    - name: "early_accuracy_10pct"
      direction: "maximize"
      description: "Test accuracy when 10% of pool labels have been acquired."
  datasets:
    - name: "breast_cancer"
      source: "sklearn.datasets.load_breast_cancer"
    - name: "wine"
      source: "sklearn.datasets.load_wine"
    - name: "digits"
      source: "sklearn.datasets.load_digits"
  compute_requirements:
    gpu_required: false
    estimated_wall_clock_sec: 600

rubric_path: "experiments/arc_bench/config/ml/rubrics/ML22.json"